Eric Mono

18, building agent systems & AI products.
Also keen on LEGO, arts and marathon.
Sharing what I build & learn in public.

[Eric Mono @EarthwalkerX17]

AI coding model can write code.

That doesn’t mean it can reliably finish software work.

I’ve released Agent OS v1.0: a local-first harness for planning, sandboxed execution, real verification, bounded recovery, and delivery.

https://t.co/CBgtslCwfx

[Eric Mono @EarthwalkerX17]

New Agent CLI demo:
one natural-language request → approved plan → 4-task DAG → parallel agents → live task view → integration → verification → acceptance → a working Retro Arcade Hub.

This is an open-source, local-first terminal agent harness I built from scratch.

[Eric Mono @EarthwalkerX17]

@DilumSanjaya Vibe coding Robotics seems a good idea. Can this be applied to real robots?

[Eric Mono @EarthwalkerX17]

@claudeai I think we need cheaper Fable more than better Opus.

[Eric Mono @EarthwalkerX17]

https://t.co/bY3WTcsBif

[Eric Mono @EarthwalkerX17]

A new end-to-end demo of Agent CLI V0.5.

Starting from a fresh Git repo and one product brief, Agent CLI built LedgerLite: a finance dashboard with charts, CSV import, persistence, tests, and a production build.

This is an open-source, local-first terminal agent harness I’m building to understand the execution, safety, state, recovery, and verification layers behind modern coding agents. GitHub in comments.

[Eric Mono @EarthwalkerX17]

Agent OS is now live on Product Hunt.

It’s an open-source, local-first AI project operating system for building, verifying, recovering, and shipping real software with specialized agents.

This is the first public v1.0.0 release. I’d really value honest feedback on the workflow and trust boundaries.

https://t.co/NZ8MwsQCIx

[Eric Mono @EarthwalkerX17]

@kevinnbass Make it exist first, then make it better.

[Eric Mono @EarthwalkerX17]

https://t.co/bY3WTcsBif

[Eric Mono @EarthwalkerX17]

Agent CLI V0.2 demo:

Starting from an empty local folder, the agent builds and tests a complete Kanban web app through natural-language interaction, then resumes the session and serves it for browser testing.

REPL, workspace trust, live tools, approvals, undo/resume—all running in one local flow.

[Eric Mono @EarthwalkerX17]

@DataChaz Why don't they just integrate everything into ONE app like Codex? Feels it's really unnecessary to classify AI roles according to human occupations.

[Eric Mono @EarthwalkerX17]

@Prathkum Excessive hype marketing often leads to one's own downfall.

[Eric Mono @EarthwalkerX17]

After finishing Agent OS, I’ve started Agent CLI: an open-source, terminal-native local agent harness.

V0.1 now has typed tools, centralized approvals, append-only logs, session resume, file snapshots/undo, evidence reports, and a verified live Anthropic loop.

The main lesson carried over from Agent OS: build a smaller kernel first, keep commits reviewable, separate approval from sandboxing, and let evidence (not summaries) prove what happened.

[Eric Mono @EarthwalkerX17]

One of the most stupid things I've ever seen done by Claude:

I asked it to make a change and create a set of clean, separated commits. It first implemented the change, then reverted everything for the first commit, and then reapplied them for the second one.

What are you doing?

[Eric Mono @EarthwalkerX17]

https://t.co/o86D5HeIsI

[Eric Mono @EarthwalkerX17]

Final demo of Agent OS.

It’s a local-first, open-source AI project operating system for planning, building, verifying, and shipping real software with agents.

This short clip gives a quick look at the full product and its core workflow. The full YouTube video is in comments.

[Eric Mono @EarthwalkerX17]

@claudeai Good strategy. Keep extending access so people rush to use it everytime.

[Eric Mono @EarthwalkerX17]

This 56s Agent OS demo was made entirely with:
Claude Code + Playwright + FFmpeg

Claude launched the local app, navigated the real UI, recorded the scenes, added captions/cards, and exported the final MP4.

No manual recording or video editing.

Claude Opus 4.8 · ~30 min · ~50K tokens.

A simple but genuinely useful workflow for quick product demos.

[Eric Mono @EarthwalkerX17]

A short demo of Agent OS:
plan → execute → verify → recover → approve → ship

Everything shown comes from a real Pulseboard run. No LLM re-run or fabricated UI.

Agent OS is the local-first harness around coding models.

[Eric Mono @EarthwalkerX17]

@EHuanglu altman's really a good actor 😂

[Eric Mono @EarthwalkerX17]

A short demo of Agent OS:
plan → execute → verify → recover → approve → ship

Everything shown comes from a real Pulseboard run. No LLM re-run or fabricated UI.

Agent OS is the local-first harness around coding models.

[Eric Mono @EarthwalkerX17]

AI coding model can write code.

That doesn’t mean it can reliably finish software work.

I’ve released Agent OS v1.0: a local-first harness for planning, sandboxed execution, real verification, bounded recovery, and delivery.

https://t.co/CBgtslCwfx

[Eric Mono @EarthwalkerX17]

@DrozdFelipe That's decided by a typed recovery matrix, targeting different failure modes like build/runtime/visual/integration error, etc.
Each type has its own contract such as accepted evidence and max attempts, so it won't stuck at repairing (most time 1~2 attempts get the work done).

[Eric Mono @EarthwalkerX17]

AI coding model can write code.

That doesn’t mean it can reliably finish software work.

I’ve released Agent OS v1.0: a local-first harness for planning, sandboxed execution, real verification, bounded recovery, and delivery.

https://t.co/CBgtslCwfx